@ls-stack/agent-eval 0.55.0 → 0.55.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/{app-CunZ8Dku.mjs → app-NI4to6lp.mjs} +87 -4
- package/dist/apps/web/dist/assets/{index-CvsPmlHl.js → index-C7QjETk8.js} +54 -54
- package/dist/apps/web/dist/index.html +1 -1
- package/dist/bin.mjs +1 -1
- package/dist/caseChild.mjs +167 -0
- package/dist/{cli-rvPrUj6S.mjs → cli-Bu9347r1.mjs} +60 -12
- package/dist/index.d.mts +151 -138
- package/dist/index.mjs +3 -3
- package/dist/runChild.mjs +3 -1
- package/dist/{runOrchestration-BWyE5lRX.mjs → runExecution-C31dpemR.mjs} +394 -1609
- package/dist/runOrchestration-3RoHLW4U.mjs +1596 -0
- package/dist/{runner-CFQ8LZmY.mjs → runner-B4EfMn1d.mjs} +2 -2
- package/dist/{runner-C2fvjKZP.mjs → runner-CTp9zHbM.mjs} +1 -1
- package/dist/{src-DEENkbkn.mjs → src-FR60ZR_4.mjs} +2 -2
- package/package.json +3 -3
- package/skills/agent-eval/SKILL.md +19 -8
package/dist/index.d.mts
CHANGED
|
@@ -1399,7 +1399,7 @@ type EvalLogLevelInput = RunLogLevel$1 | 'warning';
|
|
|
1399
1399
|
declare class EvalAssertionError extends Error {
|
|
1400
1400
|
constructor(message: string);
|
|
1401
1401
|
}
|
|
1402
|
-
/** Return the
|
|
1402
|
+
/** Return the host process clock, bypassing the eval Date shim. */
|
|
1403
1403
|
/**
|
|
1404
1404
|
* Eval time helpers for reading and moving the active eval clock.
|
|
1405
1405
|
*
|
|
@@ -1781,14 +1781,14 @@ declare const columnFormatSchema: z$1.ZodEnum<{
|
|
|
1781
1781
|
number: "number";
|
|
1782
1782
|
boolean: "boolean";
|
|
1783
1783
|
file: "file";
|
|
1784
|
-
json: "json";
|
|
1785
|
-
duration: "duration";
|
|
1786
1784
|
markdown: "markdown";
|
|
1785
|
+
json: "json";
|
|
1787
1786
|
image: "image";
|
|
1788
1787
|
html: "html";
|
|
1789
1788
|
pdf: "pdf";
|
|
1790
1789
|
audio: "audio";
|
|
1791
1790
|
video: "video";
|
|
1791
|
+
duration: "duration";
|
|
1792
1792
|
percent: "percent";
|
|
1793
1793
|
passFail: "passFail";
|
|
1794
1794
|
stars: "stars";
|
|
@@ -1808,14 +1808,14 @@ declare const columnDefSchema: z$1.ZodObject<{
|
|
|
1808
1808
|
number: "number";
|
|
1809
1809
|
boolean: "boolean";
|
|
1810
1810
|
file: "file";
|
|
1811
|
-
json: "json";
|
|
1812
|
-
duration: "duration";
|
|
1813
1811
|
markdown: "markdown";
|
|
1812
|
+
json: "json";
|
|
1814
1813
|
image: "image";
|
|
1815
1814
|
html: "html";
|
|
1816
1815
|
pdf: "pdf";
|
|
1817
1816
|
audio: "audio";
|
|
1818
1817
|
video: "video";
|
|
1818
|
+
duration: "duration";
|
|
1819
1819
|
percent: "percent";
|
|
1820
1820
|
passFail: "passFail";
|
|
1821
1821
|
stars: "stars";
|
|
@@ -1829,8 +1829,8 @@ declare const columnDefSchema: z$1.ZodObject<{
|
|
|
1829
1829
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
1830
1830
|
align: z$1.ZodOptional<z$1.ZodEnum<{
|
|
1831
1831
|
left: "left";
|
|
1832
|
-
right: "right";
|
|
1833
1832
|
center: "center";
|
|
1833
|
+
right: "right";
|
|
1834
1834
|
}>>;
|
|
1835
1835
|
}, z$1.core.$strip>;
|
|
1836
1836
|
/** Column definition exposed to the UI for eval and case tables. */
|
|
@@ -2099,10 +2099,10 @@ type EvalFreshnessStatus = z$1.infer<typeof evalFreshnessStatusSchema>;
|
|
|
2099
2099
|
* `best` selects the highest finite value and `worst` selects the lowest.
|
|
2100
2100
|
*/
|
|
2101
2101
|
declare const evalStatAggregateSchema: z$1.ZodEnum<{
|
|
2102
|
-
min: "min";
|
|
2103
|
-
max: "max";
|
|
2104
2102
|
sum: "sum";
|
|
2105
2103
|
avg: "avg";
|
|
2104
|
+
min: "min";
|
|
2105
|
+
max: "max";
|
|
2106
2106
|
best: "best";
|
|
2107
2107
|
worst: "worst";
|
|
2108
2108
|
}>;
|
|
@@ -2131,10 +2131,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2131
2131
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2132
2132
|
kind: z$1.ZodLiteral<"duration">;
|
|
2133
2133
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2134
|
-
min: "min";
|
|
2135
|
-
max: "max";
|
|
2136
2134
|
sum: "sum";
|
|
2137
2135
|
avg: "avg";
|
|
2136
|
+
min: "min";
|
|
2137
|
+
max: "max";
|
|
2138
2138
|
best: "best";
|
|
2139
2139
|
worst: "worst";
|
|
2140
2140
|
}>>;
|
|
@@ -2142,10 +2142,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2142
2142
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2143
2143
|
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2144
2144
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2145
|
-
min: "min";
|
|
2146
|
-
max: "max";
|
|
2147
2145
|
sum: "sum";
|
|
2148
2146
|
avg: "avg";
|
|
2147
|
+
min: "min";
|
|
2148
|
+
max: "max";
|
|
2149
2149
|
best: "best";
|
|
2150
2150
|
worst: "worst";
|
|
2151
2151
|
}>>;
|
|
@@ -2155,10 +2155,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2155
2155
|
key: z$1.ZodString;
|
|
2156
2156
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2157
2157
|
aggregate: z$1.ZodEnum<{
|
|
2158
|
-
min: "min";
|
|
2159
|
-
max: "max";
|
|
2160
2158
|
sum: "sum";
|
|
2161
2159
|
avg: "avg";
|
|
2160
|
+
min: "min";
|
|
2161
|
+
max: "max";
|
|
2162
2162
|
best: "best";
|
|
2163
2163
|
worst: "worst";
|
|
2164
2164
|
}>;
|
|
@@ -2166,14 +2166,14 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2166
2166
|
number: "number";
|
|
2167
2167
|
boolean: "boolean";
|
|
2168
2168
|
file: "file";
|
|
2169
|
-
json: "json";
|
|
2170
|
-
duration: "duration";
|
|
2171
2169
|
markdown: "markdown";
|
|
2170
|
+
json: "json";
|
|
2172
2171
|
image: "image";
|
|
2173
2172
|
html: "html";
|
|
2174
2173
|
pdf: "pdf";
|
|
2175
2174
|
audio: "audio";
|
|
2176
2175
|
video: "video";
|
|
2176
|
+
duration: "duration";
|
|
2177
2177
|
percent: "percent";
|
|
2178
2178
|
passFail: "passFail";
|
|
2179
2179
|
stars: "stars";
|
|
@@ -2195,10 +2195,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
|
|
|
2195
2195
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2196
2196
|
kind: z$1.ZodLiteral<"duration">;
|
|
2197
2197
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2198
|
-
min: "min";
|
|
2199
|
-
max: "max";
|
|
2200
2198
|
sum: "sum";
|
|
2201
2199
|
avg: "avg";
|
|
2200
|
+
min: "min";
|
|
2201
|
+
max: "max";
|
|
2202
2202
|
best: "best";
|
|
2203
2203
|
worst: "worst";
|
|
2204
2204
|
}>>;
|
|
@@ -2206,10 +2206,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
|
|
|
2206
2206
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2207
2207
|
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2208
2208
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2209
|
-
min: "min";
|
|
2210
|
-
max: "max";
|
|
2211
2209
|
sum: "sum";
|
|
2212
2210
|
avg: "avg";
|
|
2211
|
+
min: "min";
|
|
2212
|
+
max: "max";
|
|
2213
2213
|
best: "best";
|
|
2214
2214
|
worst: "worst";
|
|
2215
2215
|
}>>;
|
|
@@ -2219,10 +2219,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
|
|
|
2219
2219
|
key: z$1.ZodString;
|
|
2220
2220
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2221
2221
|
aggregate: z$1.ZodEnum<{
|
|
2222
|
-
min: "min";
|
|
2223
|
-
max: "max";
|
|
2224
2222
|
sum: "sum";
|
|
2225
2223
|
avg: "avg";
|
|
2224
|
+
min: "min";
|
|
2225
|
+
max: "max";
|
|
2226
2226
|
best: "best";
|
|
2227
2227
|
worst: "worst";
|
|
2228
2228
|
}>;
|
|
@@ -2230,14 +2230,14 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
|
|
|
2230
2230
|
number: "number";
|
|
2231
2231
|
boolean: "boolean";
|
|
2232
2232
|
file: "file";
|
|
2233
|
-
json: "json";
|
|
2234
|
-
duration: "duration";
|
|
2235
2233
|
markdown: "markdown";
|
|
2234
|
+
json: "json";
|
|
2236
2235
|
image: "image";
|
|
2237
2236
|
html: "html";
|
|
2238
2237
|
pdf: "pdf";
|
|
2239
2238
|
audio: "audio";
|
|
2240
2239
|
video: "video";
|
|
2240
|
+
duration: "duration";
|
|
2241
2241
|
percent: "percent";
|
|
2242
2242
|
passFail: "passFail";
|
|
2243
2243
|
stars: "stars";
|
|
@@ -2276,14 +2276,14 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2276
2276
|
number: "number";
|
|
2277
2277
|
boolean: "boolean";
|
|
2278
2278
|
file: "file";
|
|
2279
|
-
json: "json";
|
|
2280
|
-
duration: "duration";
|
|
2281
2279
|
markdown: "markdown";
|
|
2280
|
+
json: "json";
|
|
2282
2281
|
image: "image";
|
|
2283
2282
|
html: "html";
|
|
2284
2283
|
pdf: "pdf";
|
|
2285
2284
|
audio: "audio";
|
|
2286
2285
|
video: "video";
|
|
2286
|
+
duration: "duration";
|
|
2287
2287
|
percent: "percent";
|
|
2288
2288
|
passFail: "passFail";
|
|
2289
2289
|
stars: "stars";
|
|
@@ -2297,8 +2297,8 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2297
2297
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2298
2298
|
align: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2299
2299
|
left: "left";
|
|
2300
|
-
right: "right";
|
|
2301
2300
|
center: "center";
|
|
2301
|
+
right: "right";
|
|
2302
2302
|
}>>;
|
|
2303
2303
|
}, z$1.core.$strip>>;
|
|
2304
2304
|
caseCount: z$1.ZodNullable<z$1.ZodNumber>;
|
|
@@ -2322,10 +2322,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2322
2322
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2323
2323
|
kind: z$1.ZodLiteral<"duration">;
|
|
2324
2324
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2325
|
-
min: "min";
|
|
2326
|
-
max: "max";
|
|
2327
2325
|
sum: "sum";
|
|
2328
2326
|
avg: "avg";
|
|
2327
|
+
min: "min";
|
|
2328
|
+
max: "max";
|
|
2329
2329
|
best: "best";
|
|
2330
2330
|
worst: "worst";
|
|
2331
2331
|
}>>;
|
|
@@ -2333,10 +2333,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2333
2333
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2334
2334
|
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2335
2335
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2336
|
-
min: "min";
|
|
2337
|
-
max: "max";
|
|
2338
2336
|
sum: "sum";
|
|
2339
2337
|
avg: "avg";
|
|
2338
|
+
min: "min";
|
|
2339
|
+
max: "max";
|
|
2340
2340
|
best: "best";
|
|
2341
2341
|
worst: "worst";
|
|
2342
2342
|
}>>;
|
|
@@ -2346,10 +2346,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2346
2346
|
key: z$1.ZodString;
|
|
2347
2347
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2348
2348
|
aggregate: z$1.ZodEnum<{
|
|
2349
|
-
min: "min";
|
|
2350
|
-
max: "max";
|
|
2351
2349
|
sum: "sum";
|
|
2352
2350
|
avg: "avg";
|
|
2351
|
+
min: "min";
|
|
2352
|
+
max: "max";
|
|
2353
2353
|
best: "best";
|
|
2354
2354
|
worst: "worst";
|
|
2355
2355
|
}>;
|
|
@@ -2357,14 +2357,14 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2357
2357
|
number: "number";
|
|
2358
2358
|
boolean: "boolean";
|
|
2359
2359
|
file: "file";
|
|
2360
|
-
json: "json";
|
|
2361
|
-
duration: "duration";
|
|
2362
2360
|
markdown: "markdown";
|
|
2361
|
+
json: "json";
|
|
2363
2362
|
image: "image";
|
|
2364
2363
|
html: "html";
|
|
2365
2364
|
pdf: "pdf";
|
|
2366
2365
|
audio: "audio";
|
|
2367
2366
|
video: "video";
|
|
2367
|
+
duration: "duration";
|
|
2368
2368
|
percent: "percent";
|
|
2369
2369
|
passFail: "passFail";
|
|
2370
2370
|
stars: "stars";
|
|
@@ -2373,10 +2373,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2373
2373
|
accent: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2374
2374
|
}, z$1.core.$strip>], "kind">>>;
|
|
2375
2375
|
defaultStatAggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2376
|
-
min: "min";
|
|
2377
|
-
max: "max";
|
|
2378
2376
|
sum: "sum";
|
|
2379
2377
|
avg: "avg";
|
|
2378
|
+
min: "min";
|
|
2379
|
+
max: "max";
|
|
2380
2380
|
best: "best";
|
|
2381
2381
|
worst: "worst";
|
|
2382
2382
|
}>>;
|
|
@@ -2392,16 +2392,16 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2392
2392
|
metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
2393
2393
|
source: z$1.ZodLiteral<"builtin">;
|
|
2394
2394
|
metric: z$1.ZodEnum<{
|
|
2395
|
-
durationMs: "durationMs";
|
|
2396
2395
|
passRate: "passRate";
|
|
2396
|
+
durationMs: "durationMs";
|
|
2397
2397
|
}>;
|
|
2398
2398
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2399
2399
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2400
|
-
error: "error";
|
|
2401
2400
|
success: "success";
|
|
2401
|
+
error: "error";
|
|
2402
|
+
warning: "warning";
|
|
2402
2403
|
accent: "accent";
|
|
2403
2404
|
accentDim: "accentDim";
|
|
2404
|
-
warning: "warning";
|
|
2405
2405
|
textMuted: "textMuted";
|
|
2406
2406
|
}>>;
|
|
2407
2407
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -2412,20 +2412,20 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2412
2412
|
source: z$1.ZodLiteral<"column">;
|
|
2413
2413
|
key: z$1.ZodString;
|
|
2414
2414
|
aggregate: z$1.ZodEnum<{
|
|
2415
|
-
min: "min";
|
|
2416
|
-
max: "max";
|
|
2417
2415
|
sum: "sum";
|
|
2418
2416
|
avg: "avg";
|
|
2417
|
+
min: "min";
|
|
2418
|
+
max: "max";
|
|
2419
2419
|
latest: "latest";
|
|
2420
2420
|
passThresholdRate: "passThresholdRate";
|
|
2421
2421
|
}>;
|
|
2422
2422
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2423
2423
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2424
|
-
error: "error";
|
|
2425
2424
|
success: "success";
|
|
2425
|
+
error: "error";
|
|
2426
|
+
warning: "warning";
|
|
2426
2427
|
accent: "accent";
|
|
2427
2428
|
accentDim: "accentDim";
|
|
2428
|
-
warning: "warning";
|
|
2429
2429
|
textMuted: "textMuted";
|
|
2430
2430
|
}>>;
|
|
2431
2431
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -2446,18 +2446,18 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2446
2446
|
tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
2447
2447
|
source: z$1.ZodLiteral<"builtin">;
|
|
2448
2448
|
metric: z$1.ZodEnum<{
|
|
2449
|
-
durationMs: "durationMs";
|
|
2450
2449
|
passRate: "passRate";
|
|
2450
|
+
durationMs: "durationMs";
|
|
2451
2451
|
}>;
|
|
2452
2452
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2453
2453
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2454
2454
|
source: z$1.ZodLiteral<"column">;
|
|
2455
2455
|
key: z$1.ZodString;
|
|
2456
2456
|
aggregate: z$1.ZodEnum<{
|
|
2457
|
-
min: "min";
|
|
2458
|
-
max: "max";
|
|
2459
2457
|
sum: "sum";
|
|
2460
2458
|
avg: "avg";
|
|
2459
|
+
min: "min";
|
|
2460
|
+
max: "max";
|
|
2461
2461
|
latest: "latest";
|
|
2462
2462
|
passThresholdRate: "passThresholdRate";
|
|
2463
2463
|
}>;
|
|
@@ -2556,9 +2556,9 @@ declare const caseRowSchema$1: z$1.ZodObject<{
|
|
|
2556
2556
|
error: "error";
|
|
2557
2557
|
running: "running";
|
|
2558
2558
|
cancelled: "cancelled";
|
|
2559
|
-
pending: "pending";
|
|
2560
2559
|
pass: "pass";
|
|
2561
2560
|
fail: "fail";
|
|
2561
|
+
pending: "pending";
|
|
2562
2562
|
}>;
|
|
2563
2563
|
durationMs: z$1.ZodNullable<z$1.ZodNumber>;
|
|
2564
2564
|
cacheHits: z$1.ZodOptional<z$1.ZodNumber>;
|
|
@@ -2719,10 +2719,10 @@ declare const scoreTraceSchema: z$1.ZodObject<{
|
|
|
2719
2719
|
namespace: z$1.ZodString;
|
|
2720
2720
|
key: z$1.ZodString;
|
|
2721
2721
|
status: z$1.ZodEnum<{
|
|
2722
|
-
bypass: "bypass";
|
|
2723
|
-
refresh: "refresh";
|
|
2724
2722
|
hit: "hit";
|
|
2725
2723
|
miss: "miss";
|
|
2724
|
+
refresh: "refresh";
|
|
2725
|
+
bypass: "bypass";
|
|
2726
2726
|
}>;
|
|
2727
2727
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2728
2728
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -2743,9 +2743,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2743
2743
|
error: "error";
|
|
2744
2744
|
running: "running";
|
|
2745
2745
|
cancelled: "cancelled";
|
|
2746
|
-
pending: "pending";
|
|
2747
2746
|
pass: "pass";
|
|
2748
2747
|
fail: "fail";
|
|
2748
|
+
pending: "pending";
|
|
2749
2749
|
}>;
|
|
2750
2750
|
input: z$1.ZodUnknown;
|
|
2751
2751
|
trace: z$1.ZodArray<z$1.ZodObject<{
|
|
@@ -2891,10 +2891,10 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2891
2891
|
namespace: z$1.ZodString;
|
|
2892
2892
|
key: z$1.ZodString;
|
|
2893
2893
|
status: z$1.ZodEnum<{
|
|
2894
|
-
bypass: "bypass";
|
|
2895
|
-
refresh: "refresh";
|
|
2896
2894
|
hit: "hit";
|
|
2897
2895
|
miss: "miss";
|
|
2896
|
+
refresh: "refresh";
|
|
2897
|
+
bypass: "bypass";
|
|
2898
2898
|
}>;
|
|
2899
2899
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2900
2900
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -2960,10 +2960,10 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2960
2960
|
namespace: z$1.ZodString;
|
|
2961
2961
|
key: z$1.ZodString;
|
|
2962
2962
|
status: z$1.ZodEnum<{
|
|
2963
|
-
bypass: "bypass";
|
|
2964
|
-
refresh: "refresh";
|
|
2965
2963
|
hit: "hit";
|
|
2966
2964
|
miss: "miss";
|
|
2965
|
+
refresh: "refresh";
|
|
2966
|
+
bypass: "bypass";
|
|
2967
2967
|
}>;
|
|
2968
2968
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2969
2969
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -3006,8 +3006,8 @@ type EvalChartType = z$1.infer<typeof evalChartTypeSchema>;
|
|
|
3006
3006
|
* than from a per-case column.
|
|
3007
3007
|
*/
|
|
3008
3008
|
declare const evalChartBuiltinMetricSchema: z$1.ZodEnum<{
|
|
3009
|
-
durationMs: "durationMs";
|
|
3010
3009
|
passRate: "passRate";
|
|
3010
|
+
durationMs: "durationMs";
|
|
3011
3011
|
}>;
|
|
3012
3012
|
/**
|
|
3013
3013
|
* Run-level metric sourced from the aggregated `RunSummary` for a run, rather
|
|
@@ -3016,10 +3016,10 @@ declare const evalChartBuiltinMetricSchema: z$1.ZodEnum<{
|
|
|
3016
3016
|
type EvalChartBuiltinMetric = z$1.infer<typeof evalChartBuiltinMetricSchema>;
|
|
3017
3017
|
/** Reducer applied to a numeric column across all cases of a single run. */
|
|
3018
3018
|
declare const evalChartAggregateSchema: z$1.ZodEnum<{
|
|
3019
|
-
min: "min";
|
|
3020
|
-
max: "max";
|
|
3021
3019
|
sum: "sum";
|
|
3022
3020
|
avg: "avg";
|
|
3021
|
+
min: "min";
|
|
3022
|
+
max: "max";
|
|
3023
3023
|
latest: "latest";
|
|
3024
3024
|
passThresholdRate: "passThresholdRate";
|
|
3025
3025
|
}>;
|
|
@@ -3030,11 +3030,11 @@ type EvalChartAggregate = z$1.infer<typeof evalChartAggregateSchema>;
|
|
|
3030
3030
|
* not emit raw hex so authored evals stay decoupled from the web theme.
|
|
3031
3031
|
*/
|
|
3032
3032
|
declare const evalChartColorSchema: z$1.ZodEnum<{
|
|
3033
|
-
error: "error";
|
|
3034
3033
|
success: "success";
|
|
3034
|
+
error: "error";
|
|
3035
|
+
warning: "warning";
|
|
3035
3036
|
accent: "accent";
|
|
3036
3037
|
accentDim: "accentDim";
|
|
3037
|
-
warning: "warning";
|
|
3038
3038
|
textMuted: "textMuted";
|
|
3039
3039
|
}>;
|
|
3040
3040
|
/** Semantic color token resolved to a theme color by the web UI. */
|
|
@@ -3054,16 +3054,16 @@ type EvalChartAxis = z$1.infer<typeof evalChartAxisSchema>;
|
|
|
3054
3054
|
declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3055
3055
|
source: z$1.ZodLiteral<"builtin">;
|
|
3056
3056
|
metric: z$1.ZodEnum<{
|
|
3057
|
-
durationMs: "durationMs";
|
|
3058
3057
|
passRate: "passRate";
|
|
3058
|
+
durationMs: "durationMs";
|
|
3059
3059
|
}>;
|
|
3060
3060
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3061
3061
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3062
|
-
error: "error";
|
|
3063
3062
|
success: "success";
|
|
3063
|
+
error: "error";
|
|
3064
|
+
warning: "warning";
|
|
3064
3065
|
accent: "accent";
|
|
3065
3066
|
accentDim: "accentDim";
|
|
3066
|
-
warning: "warning";
|
|
3067
3067
|
textMuted: "textMuted";
|
|
3068
3068
|
}>>;
|
|
3069
3069
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3074,20 +3074,20 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
3074
3074
|
source: z$1.ZodLiteral<"column">;
|
|
3075
3075
|
key: z$1.ZodString;
|
|
3076
3076
|
aggregate: z$1.ZodEnum<{
|
|
3077
|
-
min: "min";
|
|
3078
|
-
max: "max";
|
|
3079
3077
|
sum: "sum";
|
|
3080
3078
|
avg: "avg";
|
|
3079
|
+
min: "min";
|
|
3080
|
+
max: "max";
|
|
3081
3081
|
latest: "latest";
|
|
3082
3082
|
passThresholdRate: "passThresholdRate";
|
|
3083
3083
|
}>;
|
|
3084
3084
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3085
3085
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3086
|
-
error: "error";
|
|
3087
3086
|
success: "success";
|
|
3087
|
+
error: "error";
|
|
3088
|
+
warning: "warning";
|
|
3088
3089
|
accent: "accent";
|
|
3089
3090
|
accentDim: "accentDim";
|
|
3090
|
-
warning: "warning";
|
|
3091
3091
|
textMuted: "textMuted";
|
|
3092
3092
|
}>>;
|
|
3093
3093
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3101,18 +3101,18 @@ type EvalChartMetric = z$1.infer<typeof evalChartMetricSchema>;
|
|
|
3101
3101
|
declare const evalChartTooltipExtraSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3102
3102
|
source: z$1.ZodLiteral<"builtin">;
|
|
3103
3103
|
metric: z$1.ZodEnum<{
|
|
3104
|
-
durationMs: "durationMs";
|
|
3105
3104
|
passRate: "passRate";
|
|
3105
|
+
durationMs: "durationMs";
|
|
3106
3106
|
}>;
|
|
3107
3107
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3108
3108
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
3109
3109
|
source: z$1.ZodLiteral<"column">;
|
|
3110
3110
|
key: z$1.ZodString;
|
|
3111
3111
|
aggregate: z$1.ZodEnum<{
|
|
3112
|
-
min: "min";
|
|
3113
|
-
max: "max";
|
|
3114
3112
|
sum: "sum";
|
|
3115
3113
|
avg: "avg";
|
|
3114
|
+
min: "min";
|
|
3115
|
+
max: "max";
|
|
3116
3116
|
latest: "latest";
|
|
3117
3117
|
passThresholdRate: "passThresholdRate";
|
|
3118
3118
|
}>;
|
|
@@ -3137,16 +3137,16 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
3137
3137
|
metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3138
3138
|
source: z$1.ZodLiteral<"builtin">;
|
|
3139
3139
|
metric: z$1.ZodEnum<{
|
|
3140
|
-
durationMs: "durationMs";
|
|
3141
3140
|
passRate: "passRate";
|
|
3141
|
+
durationMs: "durationMs";
|
|
3142
3142
|
}>;
|
|
3143
3143
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3144
3144
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3145
|
-
error: "error";
|
|
3146
3145
|
success: "success";
|
|
3146
|
+
error: "error";
|
|
3147
|
+
warning: "warning";
|
|
3147
3148
|
accent: "accent";
|
|
3148
3149
|
accentDim: "accentDim";
|
|
3149
|
-
warning: "warning";
|
|
3150
3150
|
textMuted: "textMuted";
|
|
3151
3151
|
}>>;
|
|
3152
3152
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3157,20 +3157,20 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
3157
3157
|
source: z$1.ZodLiteral<"column">;
|
|
3158
3158
|
key: z$1.ZodString;
|
|
3159
3159
|
aggregate: z$1.ZodEnum<{
|
|
3160
|
-
min: "min";
|
|
3161
|
-
max: "max";
|
|
3162
3160
|
sum: "sum";
|
|
3163
3161
|
avg: "avg";
|
|
3162
|
+
min: "min";
|
|
3163
|
+
max: "max";
|
|
3164
3164
|
latest: "latest";
|
|
3165
3165
|
passThresholdRate: "passThresholdRate";
|
|
3166
3166
|
}>;
|
|
3167
3167
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3168
3168
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3169
|
-
error: "error";
|
|
3170
3169
|
success: "success";
|
|
3170
|
+
error: "error";
|
|
3171
|
+
warning: "warning";
|
|
3171
3172
|
accent: "accent";
|
|
3172
3173
|
accentDim: "accentDim";
|
|
3173
|
-
warning: "warning";
|
|
3174
3174
|
textMuted: "textMuted";
|
|
3175
3175
|
}>>;
|
|
3176
3176
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3191,18 +3191,18 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
3191
3191
|
tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3192
3192
|
source: z$1.ZodLiteral<"builtin">;
|
|
3193
3193
|
metric: z$1.ZodEnum<{
|
|
3194
|
-
durationMs: "durationMs";
|
|
3195
3194
|
passRate: "passRate";
|
|
3195
|
+
durationMs: "durationMs";
|
|
3196
3196
|
}>;
|
|
3197
3197
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3198
3198
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
3199
3199
|
source: z$1.ZodLiteral<"column">;
|
|
3200
3200
|
key: z$1.ZodString;
|
|
3201
3201
|
aggregate: z$1.ZodEnum<{
|
|
3202
|
-
min: "min";
|
|
3203
|
-
max: "max";
|
|
3204
3202
|
sum: "sum";
|
|
3205
3203
|
avg: "avg";
|
|
3204
|
+
min: "min";
|
|
3205
|
+
max: "max";
|
|
3206
3206
|
latest: "latest";
|
|
3207
3207
|
passThresholdRate: "passThresholdRate";
|
|
3208
3208
|
}>;
|
|
@@ -3227,16 +3227,16 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3227
3227
|
metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3228
3228
|
source: z$1.ZodLiteral<"builtin">;
|
|
3229
3229
|
metric: z$1.ZodEnum<{
|
|
3230
|
-
durationMs: "durationMs";
|
|
3231
3230
|
passRate: "passRate";
|
|
3231
|
+
durationMs: "durationMs";
|
|
3232
3232
|
}>;
|
|
3233
3233
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3234
3234
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3235
|
-
error: "error";
|
|
3236
3235
|
success: "success";
|
|
3236
|
+
error: "error";
|
|
3237
|
+
warning: "warning";
|
|
3237
3238
|
accent: "accent";
|
|
3238
3239
|
accentDim: "accentDim";
|
|
3239
|
-
warning: "warning";
|
|
3240
3240
|
textMuted: "textMuted";
|
|
3241
3241
|
}>>;
|
|
3242
3242
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3247,20 +3247,20 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3247
3247
|
source: z$1.ZodLiteral<"column">;
|
|
3248
3248
|
key: z$1.ZodString;
|
|
3249
3249
|
aggregate: z$1.ZodEnum<{
|
|
3250
|
-
min: "min";
|
|
3251
|
-
max: "max";
|
|
3252
3250
|
sum: "sum";
|
|
3253
3251
|
avg: "avg";
|
|
3252
|
+
min: "min";
|
|
3253
|
+
max: "max";
|
|
3254
3254
|
latest: "latest";
|
|
3255
3255
|
passThresholdRate: "passThresholdRate";
|
|
3256
3256
|
}>;
|
|
3257
3257
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3258
3258
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3259
|
-
error: "error";
|
|
3260
3259
|
success: "success";
|
|
3260
|
+
error: "error";
|
|
3261
|
+
warning: "warning";
|
|
3261
3262
|
accent: "accent";
|
|
3262
3263
|
accentDim: "accentDim";
|
|
3263
|
-
warning: "warning";
|
|
3264
3264
|
textMuted: "textMuted";
|
|
3265
3265
|
}>>;
|
|
3266
3266
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3281,18 +3281,18 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3281
3281
|
tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3282
3282
|
source: z$1.ZodLiteral<"builtin">;
|
|
3283
3283
|
metric: z$1.ZodEnum<{
|
|
3284
|
-
durationMs: "durationMs";
|
|
3285
3284
|
passRate: "passRate";
|
|
3285
|
+
durationMs: "durationMs";
|
|
3286
3286
|
}>;
|
|
3287
3287
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3288
3288
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
3289
3289
|
source: z$1.ZodLiteral<"column">;
|
|
3290
3290
|
key: z$1.ZodString;
|
|
3291
3291
|
aggregate: z$1.ZodEnum<{
|
|
3292
|
-
min: "min";
|
|
3293
|
-
max: "max";
|
|
3294
3292
|
sum: "sum";
|
|
3295
3293
|
avg: "avg";
|
|
3294
|
+
min: "min";
|
|
3295
|
+
max: "max";
|
|
3296
3296
|
latest: "latest";
|
|
3297
3297
|
passThresholdRate: "passThresholdRate";
|
|
3298
3298
|
}>;
|
|
@@ -3321,8 +3321,8 @@ declare const runManifestSchema$1: z$1.ZodObject<{
|
|
|
3321
3321
|
target: z$1.ZodObject<{
|
|
3322
3322
|
mode: z$1.ZodEnum<{
|
|
3323
3323
|
all: "all";
|
|
3324
|
-
evalIds: "evalIds";
|
|
3325
3324
|
caseIds: "caseIds";
|
|
3325
|
+
evalIds: "evalIds";
|
|
3326
3326
|
}>;
|
|
3327
3327
|
evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
3328
3328
|
files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
@@ -3336,9 +3336,9 @@ declare const runManifestSchema$1: z$1.ZodObject<{
|
|
|
3336
3336
|
median: "median";
|
|
3337
3337
|
}>>>;
|
|
3338
3338
|
cacheMode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3339
|
-
use: "use";
|
|
3340
|
-
bypass: "bypass";
|
|
3341
3339
|
refresh: "refresh";
|
|
3340
|
+
bypass: "bypass";
|
|
3341
|
+
use: "use";
|
|
3342
3342
|
}>>;
|
|
3343
3343
|
}, z$1.core.$strip>;
|
|
3344
3344
|
/** Persisted lifecycle metadata for a single eval run. */
|
|
@@ -3434,8 +3434,8 @@ type TrialSelectionMode = z$1.infer<typeof trialSelectionModeSchema>;
|
|
|
3434
3434
|
/** Built-in eval-level output/column keys. */
|
|
3435
3435
|
/** Removal config for built-in eval-level outputs and UI metadata. */
|
|
3436
3436
|
declare const removeDefaultConfigSchema: z$1.ZodUnion<readonly [z$1.ZodLiteral<true>, z$1.ZodArray<z$1.ZodEnum<{
|
|
3437
|
-
apiCalls: "apiCalls";
|
|
3438
3437
|
costUsd: "costUsd";
|
|
3438
|
+
apiCalls: "apiCalls";
|
|
3439
3439
|
llmTurns: "llmTurns";
|
|
3440
3440
|
inputTokens: "inputTokens";
|
|
3441
3441
|
outputTokens: "outputTokens";
|
|
@@ -4042,7 +4042,13 @@ type AgentEvalsConfig$1 = {
|
|
|
4042
4042
|
* Exact namespace-specific retention caps. Values override
|
|
4043
4043
|
* `maxEntriesPerNamespace` for matching namespaces.
|
|
4044
4044
|
*/
|
|
4045
|
-
maxEntriesByNamespace?: Record<string, number>;
|
|
4045
|
+
maxEntriesByNamespace?: Record<string, number>;
|
|
4046
|
+
/**
|
|
4047
|
+
* Milliseconds the runner waits after becoming idle before pruning indexed
|
|
4048
|
+
* cache entries. Defaults to `5000`; non-positive or non-finite values use
|
|
4049
|
+
* the default.
|
|
4050
|
+
*/
|
|
4051
|
+
pruneIdleDelayMs?: number; /** Legacy alias for `maxEntriesPerNamespace`, retained so older config files keep working. */
|
|
4046
4052
|
maxEntriesPerEval?: number;
|
|
4047
4053
|
};
|
|
4048
4054
|
};
|
|
@@ -4249,9 +4255,9 @@ declare function extractApiCalls(spans: EvalTraceSpan$1[], config: ResolvedApiCa
|
|
|
4249
4255
|
* - `refresh`: never read, always write (forces re-execution and overwrites).
|
|
4250
4256
|
*/
|
|
4251
4257
|
declare const cacheModeSchema: z$1.ZodEnum<{
|
|
4252
|
-
use: "use";
|
|
4253
|
-
bypass: "bypass";
|
|
4254
4258
|
refresh: "refresh";
|
|
4259
|
+
bypass: "bypass";
|
|
4260
|
+
use: "use";
|
|
4255
4261
|
}>;
|
|
4256
4262
|
/** Mode controlling how cached spans behave during a run. */
|
|
4257
4263
|
type CacheMode = z$1.infer<typeof cacheModeSchema>;
|
|
@@ -4265,17 +4271,17 @@ declare const spanCacheOptionsSchema: z$1.ZodObject<{
|
|
|
4265
4271
|
type SpanCacheOptions = z$1.infer<typeof spanCacheOptionsSchema>;
|
|
4266
4272
|
/** Category of operation stored in the eval cache. */
|
|
4267
4273
|
declare const cacheOperationTypeSchema: z$1.ZodEnum<{
|
|
4268
|
-
span: "span";
|
|
4269
4274
|
value: "value";
|
|
4275
|
+
span: "span";
|
|
4270
4276
|
}>;
|
|
4271
4277
|
/** Category of operation stored in the eval cache. */
|
|
4272
4278
|
type CacheOperationType = z$1.infer<typeof cacheOperationTypeSchema>;
|
|
4273
4279
|
/** Status of a cache lookup recorded on a span or case scope. */
|
|
4274
4280
|
declare const cacheStatusSchema: z$1.ZodEnum<{
|
|
4275
|
-
bypass: "bypass";
|
|
4276
|
-
refresh: "refresh";
|
|
4277
4281
|
hit: "hit";
|
|
4278
4282
|
miss: "miss";
|
|
4283
|
+
refresh: "refresh";
|
|
4284
|
+
bypass: "bypass";
|
|
4279
4285
|
}>;
|
|
4280
4286
|
/** Status of a cache lookup recorded on a span or case scope. */
|
|
4281
4287
|
type CacheStatus = z$1.infer<typeof cacheStatusSchema>;
|
|
@@ -4292,10 +4298,10 @@ declare const traceCacheRefSchema: z$1.ZodObject<{
|
|
|
4292
4298
|
namespace: z$1.ZodString;
|
|
4293
4299
|
key: z$1.ZodString;
|
|
4294
4300
|
status: z$1.ZodEnum<{
|
|
4295
|
-
bypass: "bypass";
|
|
4296
|
-
refresh: "refresh";
|
|
4297
4301
|
hit: "hit";
|
|
4298
4302
|
miss: "miss";
|
|
4303
|
+
refresh: "refresh";
|
|
4304
|
+
bypass: "bypass";
|
|
4299
4305
|
}>;
|
|
4300
4306
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
4301
4307
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -4304,22 +4310,25 @@ declare const traceCacheRefSchema: z$1.ZodObject<{
|
|
|
4304
4310
|
}, z$1.core.$strip>;
|
|
4305
4311
|
/** Reference to a value-cache lookup performed via `evalTracer.cache(...)`. */
|
|
4306
4312
|
type TraceCacheRef$1 = z$1.infer<typeof traceCacheRefSchema>;
|
|
4307
|
-
/**
|
|
4313
|
+
/** Minimal index-backed summary of a persisted cache entry. */
|
|
4308
4314
|
declare const cacheListItemSchema$1: z$1.ZodObject<{
|
|
4309
4315
|
key: z$1.ZodString;
|
|
4310
4316
|
namespace: z$1.ZodString;
|
|
4311
|
-
operationType: z$1.ZodEnum<{
|
|
4312
|
-
span: "span";
|
|
4313
|
-
value: "value";
|
|
4314
|
-
}>;
|
|
4315
|
-
operationName: z$1.ZodString;
|
|
4316
|
-
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
4317
|
-
spanKind: z$1.ZodOptional<z$1.ZodString>;
|
|
4318
4317
|
storedAt: z$1.ZodString;
|
|
4319
|
-
|
|
4318
|
+
lastAccessedAt: z$1.ZodString;
|
|
4320
4319
|
}, z$1.core.$strip>;
|
|
4321
|
-
/**
|
|
4320
|
+
/** Minimal summary row for a single cache entry. */
|
|
4322
4321
|
type CacheListItem = z$1.infer<typeof cacheListItemSchema$1>;
|
|
4322
|
+
/** Summary of cleanup performed by manual cache repair. */
|
|
4323
|
+
declare const cacheRepairSummarySchema$1: z$1.ZodObject<{
|
|
4324
|
+
removedCacheFiles: z$1.ZodNumber;
|
|
4325
|
+
removedDebugFiles: z$1.ZodNumber;
|
|
4326
|
+
removedBlobFiles: z$1.ZodNumber;
|
|
4327
|
+
removedIndexRows: z$1.ZodNumber;
|
|
4328
|
+
rewrittenIndexes: z$1.ZodNumber;
|
|
4329
|
+
}, z$1.core.$strip>;
|
|
4330
|
+
/** Stable JSON summary returned by manual cache repair. */
|
|
4331
|
+
type CacheRepairSummary = z$1.infer<typeof cacheRepairSummarySchema$1>;
|
|
4323
4332
|
/** Serialized nested span captured while recording a cached operation. */
|
|
4324
4333
|
type SerializedCacheSpan = {
|
|
4325
4334
|
kind: string;
|
|
@@ -4432,8 +4441,8 @@ declare const cacheEntrySchema: z$1.ZodObject<{
|
|
|
4432
4441
|
key: z$1.ZodString;
|
|
4433
4442
|
namespace: z$1.ZodString;
|
|
4434
4443
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4435
|
-
span: "span";
|
|
4436
4444
|
value: "value";
|
|
4445
|
+
span: "span";
|
|
4437
4446
|
}>>;
|
|
4438
4447
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4439
4448
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4511,8 +4520,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4511
4520
|
key: z$1.ZodString;
|
|
4512
4521
|
namespace: z$1.ZodString;
|
|
4513
4522
|
operationType: z$1.ZodEnum<{
|
|
4514
|
-
span: "span";
|
|
4515
4523
|
value: "value";
|
|
4524
|
+
span: "span";
|
|
4516
4525
|
}>;
|
|
4517
4526
|
operationName: z$1.ZodString;
|
|
4518
4527
|
storedAt: z$1.ZodString;
|
|
@@ -4522,8 +4531,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4522
4531
|
key: z$1.ZodString;
|
|
4523
4532
|
namespace: z$1.ZodString;
|
|
4524
4533
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4525
|
-
span: "span";
|
|
4526
4534
|
value: "value";
|
|
4535
|
+
span: "span";
|
|
4527
4536
|
}>>;
|
|
4528
4537
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4529
4538
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4601,8 +4610,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4601
4610
|
key: z$1.ZodString;
|
|
4602
4611
|
namespace: z$1.ZodString;
|
|
4603
4612
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4604
|
-
span: "span";
|
|
4605
4613
|
value: "value";
|
|
4614
|
+
span: "span";
|
|
4606
4615
|
}>>;
|
|
4607
4616
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4608
4617
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4671,8 +4680,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4671
4680
|
key: z$1.ZodString;
|
|
4672
4681
|
namespace: z$1.ZodString;
|
|
4673
4682
|
operationType: z$1.ZodEnum<{
|
|
4674
|
-
span: "span";
|
|
4675
4683
|
value: "value";
|
|
4684
|
+
span: "span";
|
|
4676
4685
|
}>;
|
|
4677
4686
|
operationName: z$1.ZodString;
|
|
4678
4687
|
storedAt: z$1.ZodString;
|
|
@@ -4682,8 +4691,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4682
4691
|
key: z$1.ZodString;
|
|
4683
4692
|
namespace: z$1.ZodString;
|
|
4684
4693
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4685
|
-
span: "span";
|
|
4686
4694
|
value: "value";
|
|
4695
|
+
span: "span";
|
|
4687
4696
|
}>>;
|
|
4688
4697
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4689
4698
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4761,8 +4770,8 @@ declare const cacheFileSchema: z$1.ZodObject<{
|
|
|
4761
4770
|
key: z$1.ZodString;
|
|
4762
4771
|
namespace: z$1.ZodString;
|
|
4763
4772
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4764
|
-
span: "span";
|
|
4765
4773
|
value: "value";
|
|
4774
|
+
span: "span";
|
|
4766
4775
|
}>>;
|
|
4767
4776
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4768
4777
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4839,8 +4848,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
4839
4848
|
key: z$1.ZodString;
|
|
4840
4849
|
namespace: z$1.ZodString;
|
|
4841
4850
|
operationType: z$1.ZodEnum<{
|
|
4842
|
-
span: "span";
|
|
4843
4851
|
value: "value";
|
|
4852
|
+
span: "span";
|
|
4844
4853
|
}>;
|
|
4845
4854
|
operationName: z$1.ZodString;
|
|
4846
4855
|
storedAt: z$1.ZodString;
|
|
@@ -4850,8 +4859,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
4850
4859
|
key: z$1.ZodString;
|
|
4851
4860
|
namespace: z$1.ZodString;
|
|
4852
4861
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4853
|
-
span: "span";
|
|
4854
4862
|
value: "value";
|
|
4863
|
+
span: "span";
|
|
4855
4864
|
}>>;
|
|
4856
4865
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4857
4866
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4997,8 +5006,8 @@ type SseEnvelope = z$1.infer<typeof sseEnvelopeSchema$1>; //#endregion
|
|
|
4997
5006
|
//#region src/schemas/api.d.ts
|
|
4998
5007
|
/** Lifecycle state for an app config reload triggered by `agent-evals.config.ts`. */
|
|
4999
5008
|
declare const configReloadStatusSchema: z$1.ZodEnum<{
|
|
5000
|
-
idle: "idle";
|
|
5001
5009
|
pending: "pending";
|
|
5010
|
+
idle: "idle";
|
|
5002
5011
|
reloading: "reloading";
|
|
5003
5012
|
}>;
|
|
5004
5013
|
/** Status for config reloads in the long-running app server. */
|
|
@@ -5006,8 +5015,8 @@ type ConfigReloadStatus = z$1.infer<typeof configReloadStatusSchema>;
|
|
|
5006
5015
|
/** UI/API-visible state for config reloads in `agent-evals app`. */
|
|
5007
5016
|
declare const configReloadStateSchema$1: z$1.ZodObject<{
|
|
5008
5017
|
status: z$1.ZodEnum<{
|
|
5009
|
-
idle: "idle";
|
|
5010
5018
|
pending: "pending";
|
|
5019
|
+
idle: "idle";
|
|
5011
5020
|
reloading: "reloading";
|
|
5012
5021
|
}>;
|
|
5013
5022
|
activeRunCount: z$1.ZodNumber;
|
|
@@ -5021,8 +5030,8 @@ declare const createRunRequestSchema$1: z$1.ZodObject<{
|
|
|
5021
5030
|
target: z$1.ZodObject<{
|
|
5022
5031
|
mode: z$1.ZodEnum<{
|
|
5023
5032
|
all: "all";
|
|
5024
|
-
evalIds: "evalIds";
|
|
5025
5033
|
caseIds: "caseIds";
|
|
5034
|
+
evalIds: "evalIds";
|
|
5026
5035
|
}>;
|
|
5027
5036
|
evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
5028
5037
|
files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
@@ -5034,9 +5043,9 @@ declare const createRunRequestSchema$1: z$1.ZodObject<{
|
|
|
5034
5043
|
temporary: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
5035
5044
|
cache: z$1.ZodOptional<z$1.ZodObject<{
|
|
5036
5045
|
mode: z$1.ZodDefault<z$1.ZodEnum<{
|
|
5037
|
-
use: "use";
|
|
5038
|
-
bypass: "bypass";
|
|
5039
5046
|
refresh: "refresh";
|
|
5047
|
+
bypass: "bypass";
|
|
5048
|
+
use: "use";
|
|
5040
5049
|
}>>;
|
|
5041
5050
|
}, z$1.core.$strip>>;
|
|
5042
5051
|
manualInputs: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
@@ -6097,22 +6106,25 @@ type ResolvedLlmCallCostCurrency$1 = {
|
|
|
6097
6106
|
numberFormat?: NumberDisplayOptions$1;
|
|
6098
6107
|
}; //#endregion
|
|
6099
6108
|
//#region ../shared/src/schemas/cache.d.ts
|
|
6100
|
-
/**
|
|
6109
|
+
/** Minimal index-backed summary of a persisted cache entry. */
|
|
6101
6110
|
declare const cacheListItemSchema: z$1.ZodObject<{
|
|
6102
6111
|
key: z$1.ZodString;
|
|
6103
6112
|
namespace: z$1.ZodString;
|
|
6104
|
-
operationType: z$1.ZodEnum<{
|
|
6105
|
-
span: "span";
|
|
6106
|
-
value: "value";
|
|
6107
|
-
}>;
|
|
6108
|
-
operationName: z$1.ZodString;
|
|
6109
|
-
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
6110
|
-
spanKind: z$1.ZodOptional<z$1.ZodString>;
|
|
6111
6113
|
storedAt: z$1.ZodString;
|
|
6112
|
-
|
|
6114
|
+
lastAccessedAt: z$1.ZodString;
|
|
6113
6115
|
}, z$1.core.$strip>;
|
|
6114
|
-
/**
|
|
6116
|
+
/** Minimal summary row for a single cache entry. */
|
|
6115
6117
|
type CacheListItem$1 = z$1.infer<typeof cacheListItemSchema>;
|
|
6118
|
+
/** Summary of cleanup performed by manual cache repair. */
|
|
6119
|
+
declare const cacheRepairSummarySchema: z$1.ZodObject<{
|
|
6120
|
+
removedCacheFiles: z$1.ZodNumber;
|
|
6121
|
+
removedDebugFiles: z$1.ZodNumber;
|
|
6122
|
+
removedBlobFiles: z$1.ZodNumber;
|
|
6123
|
+
removedIndexRows: z$1.ZodNumber;
|
|
6124
|
+
rewrittenIndexes: z$1.ZodNumber;
|
|
6125
|
+
}, z$1.core.$strip>;
|
|
6126
|
+
/** Stable JSON summary returned by manual cache repair. */
|
|
6127
|
+
type CacheRepairSummary$1 = z$1.infer<typeof cacheRepairSummarySchema>;
|
|
6116
6128
|
/** Serialized nested span captured while recording a cached operation. */
|
|
6117
6129
|
type SerializedCacheSpan$1 = {
|
|
6118
6130
|
kind: string;
|
|
@@ -6454,7 +6466,8 @@ type EvalRunner = {
|
|
|
6454
6466
|
* Remove cache entries matching `filter`, or all entries when no filter is
|
|
6455
6467
|
* supplied.
|
|
6456
6468
|
*/
|
|
6457
|
-
clearCache(filter?: CacheClearFilter): Promise<void>;
|
|
6469
|
+
clearCache(filter?: CacheClearFilter): Promise<void>; /** Remove cache/debug/blob files that are not referenced by cache indexes. */
|
|
6470
|
+
repairCache(): Promise<CacheRepairSummary$1>;
|
|
6458
6471
|
/**
|
|
6459
6472
|
* Recompute persisted case and run statuses for terminal runs touching one
|
|
6460
6473
|
* eval. Accepts the exact eval key.
|
|
@@ -6635,4 +6648,4 @@ declare function defineEval<TInput = unknown, TOutputs extends EvalOutputs = Eva
|
|
|
6635
6648
|
/** Return whether the active eval case has tags matching the typed input. */
|
|
6636
6649
|
declare function matchesEvalTags(input: EvalTagMatchInput): boolean;
|
|
6637
6650
|
//#endregion
|
|
6638
|
-
export { AgentEvalTagRegistry, AgentEvalsConfig, type ApiCallEntry, type ApiCallMetric, type ApiCallMetricFormat, type ApiCallMetricPlacement, type ApiCallMetricValue, type ApiCallsConfigInput, type AssertionFailure, type CacheActivityEntry, type CacheAdapter, type CacheDebugKeyEntry, type CacheDebugKeyFile, type CacheDebugKeyWrite, type CacheEntry, type CacheEntryWithDebugKey, type CacheFile, type CacheHitEntry, type CacheKeyHashInput, type CacheKeyHashOptions, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CacheSerializationOptions, type CacheStatus, type CallDerivedAttribute, type CallDerivedAttributeContext, type CallDerivedAttributesConfig, type CallDerivedAttributesFn, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type ConfigReloadState, type ConfigReloadStatus, type CreateRunRequest, type DefaultConfigKey, type DerivedStatus, type DiscoveryIssue, EvalAssertionError, type EvalCacheConfig, EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, EvalDefinition, type EvalDeriveConfig, type EvalDeriveContext, type EvalDeriveFn, type EvalDeriveMap, type EvalDeriveValueFn, type EvalDisplayStatus, type EvalExecuteContext, type EvalExpectation, type EvalFreshnessStatus, type EvalManualInputConfig, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStartTime, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, EvalTag, EvalTagMatchInput, type EvalTraceTree, type JsonCell, type LlmCallCostBreakdown, type LlmCallCostCurrency, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallPricing, type LlmCallPricingRate, type LlmCallPricingRegistry, type LlmCallSimulatedTokens, type LlmCallsConfigInput, type LlmCostScenario, type ManualInputDescriptor, type ManualInputFieldDescriptor, type ManualInputFieldKind, type ManualInputFieldOverride, type ManualInputFieldsConfig, type ManualInputFileValue, type ManualInputSelectOption, type MaterializeManualInputFilesResult, type NumberDisplayOptions, type ReadManualInputFileResult, type RemoveDefaultConfig, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedCallDerivedAttribute, type ResolvedLlmCallCostCurrency, type ResolvedLlmCallMetric, type ResolvedLlmCallPricing, type ResolvedLlmCallsConfig, type RunInEvalScopeOptions, type RunLogEntry, type RunLogLevel, type RunLogLocation, type RunLogPhase, type RunLogsConfigInput, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SerializedCacheValue, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, appendToEvalOutput, buildTraceTree, captureEvalSpanError, cleanupStagedManualInputFiles, createRunner, defineEval, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalExpect, evalLog, evalSpan, evalTime, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, getCurrentScope, getEvalCaseInput, getEvalRegistry, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, isManualInputFileValue, manualInputFileValueSchema, matchesEvalTags, materializeManualInputFiles, mergeEvalOutput, nextEvalId, readManualInputFile, repoFile, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, serializeCacheRecording, serializeCacheValue, setEvalOutput, setScopeCacheContext, simulateLlmCallCost, simulateTokenAllocation, stageManualInputFile, stageManualInputFileFromPath, startEvalBackgroundJob, z };
|
|
6651
|
+
export { AgentEvalTagRegistry, AgentEvalsConfig, type ApiCallEntry, type ApiCallMetric, type ApiCallMetricFormat, type ApiCallMetricPlacement, type ApiCallMetricValue, type ApiCallsConfigInput, type AssertionFailure, type CacheActivityEntry, type CacheAdapter, type CacheDebugKeyEntry, type CacheDebugKeyFile, type CacheDebugKeyWrite, type CacheEntry, type CacheEntryWithDebugKey, type CacheFile, type CacheHitEntry, type CacheKeyHashInput, type CacheKeyHashOptions, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheRepairSummary, type CacheScopeContext, type CacheSerializationOptions, type CacheStatus, type CallDerivedAttribute, type CallDerivedAttributeContext, type CallDerivedAttributesConfig, type CallDerivedAttributesFn, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type ConfigReloadState, type ConfigReloadStatus, type CreateRunRequest, type DefaultConfigKey, type DerivedStatus, type DiscoveryIssue, EvalAssertionError, type EvalCacheConfig, EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, EvalDefinition, type EvalDeriveConfig, type EvalDeriveContext, type EvalDeriveFn, type EvalDeriveMap, type EvalDeriveValueFn, type EvalDisplayStatus, type EvalExecuteContext, type EvalExpectation, type EvalFreshnessStatus, type EvalManualInputConfig, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStartTime, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, EvalTag, EvalTagMatchInput, type EvalTraceTree, type JsonCell, type LlmCallCostBreakdown, type LlmCallCostCurrency, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallPricing, type LlmCallPricingRate, type LlmCallPricingRegistry, type LlmCallSimulatedTokens, type LlmCallsConfigInput, type LlmCostScenario, type ManualInputDescriptor, type ManualInputFieldDescriptor, type ManualInputFieldKind, type ManualInputFieldOverride, type ManualInputFieldsConfig, type ManualInputFileValue, type ManualInputSelectOption, type MaterializeManualInputFilesResult, type NumberDisplayOptions, type ReadManualInputFileResult, type RemoveDefaultConfig, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedCallDerivedAttribute, type ResolvedLlmCallCostCurrency, type ResolvedLlmCallMetric, type ResolvedLlmCallPricing, type ResolvedLlmCallsConfig, type RunInEvalScopeOptions, type RunLogEntry, type RunLogLevel, type RunLogLocation, type RunLogPhase, type RunLogsConfigInput, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SerializedCacheValue, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, appendToEvalOutput, buildTraceTree, captureEvalSpanError, cleanupStagedManualInputFiles, createRunner, defineEval, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalExpect, evalLog, evalSpan, evalTime, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, getCurrentScope, getEvalCaseInput, getEvalRegistry, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, isManualInputFileValue, manualInputFileValueSchema, matchesEvalTags, materializeManualInputFiles, mergeEvalOutput, nextEvalId, readManualInputFile, repoFile, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, serializeCacheRecording, serializeCacheValue, setEvalOutput, setScopeCacheContext, simulateLlmCallCost, simulateTokenAllocation, stageManualInputFile, stageManualInputFileFromPath, startEvalBackgroundJob, z };
|